A content pipeline where the editor still has the last word
Our blog posts are drafted by agents and published by a human. A walk through the writers room, the line editor on a different model family, and why the approve button is not automated.
Most "AI content" is one prompt away from the publish button. Ours takes a longer road on purpose. This is the pipeline that produces posts for our ventures, as it actually runs.
The stages
Research. A trend researcher works from tracked sources and writes inspiration packets. A deep researcher builds story packs — named facts, numbers, quotes with citations — for accepted pitches. Nothing downstream may use a specific that does not trace to a story pack.
The writers room. A senior writer model runs a pitch session with junior subagents exploring different angles. The session produces exactly five pitches. An editor pass kills the weak ones, promotes the strong ones, and parks the maybes with a resurface condition.
Drafting. A writer agent turns one ready brief into a draft. It refuses to draft when three drafts already await review — a work-in-progress limit, because a pile of unreviewed drafts is inventory, not progress.
Line editing. A separate line editor reviews every draft against our voice file. It deliberately runs on a different model family than the writer, for the same reason you don't let an author proofread their own manuscript: shared blind spots.
Publishing. The editor's verdict can request revisions, escalate to a human decision after two revision rounds, or mark the draft approved — which still only stages it. A person pushes publish.
What we learned running it
- The bottleneck should be judgment, not typing. The pipeline removed the typing. The judgment calls — is this pitch worth a reader's time? — stayed human, and got more attention because they were the only calls left.
- Different model families disagree usefully. The line editor catches voice drift the writer's own family reliably misses.
- Receipts beat memory. Every stage writes files. When a post underperforms, we can replay exactly how it got made.
The same architecture — draft with agents, gate with a named human, log everything — is what we install for clients through our AI automation service. It is not novel. It is just built all the way to the end.